libvm/mm-cand-aim_on_task_arithmetic__calib_instruction

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:May 19, 2026Architecture:Transformer Warm

The libvm/mm-cand-aim_on_task_arithmetic__calib_instruction model is an 8 billion parameter language model derived from a merge of Qwen3-8B-Base, Qwen3-8B, OpenDataArena/Qwen3-8B-ODA-Math-460k, and mlabonne/Qwen3-8B-abliterated using task arithmetic. It was further refined with AIM calibration on examples from allenai/WildChat, specializing it for improved performance on arithmetic tasks. With a 32768 token context length, this model is optimized for robust mathematical reasoning and instruction following.

Loading preview...

Model Overview

This model, libvm/mm-cand-aim_on_task_arithmetic__calib_instruction, is an 8 billion parameter language model built upon the Qwen3-8B architecture. It was created through a multi-stage process designed to enhance its arithmetic and instruction-following capabilities.

Key Capabilities

  • Arithmetic Task Performance: The model's foundation was established by merging several Qwen3-8B variants, including OpenDataArena/Qwen3-8B-ODA-Math-460k, specifically targeting mathematical reasoning.
  • Instruction Following: Further refinement involved applying AIM (Aligned Instruction Model) calibration using examples from allenai/WildChat, which helps improve its ability to follow complex instructions.
  • Merged Architecture: It combines the strengths of Qwen/Qwen3-8B-Base, Qwen/Qwen3-8B, OpenDataArena/Qwen3-8B-ODA-Math-460k, and mlabonne/Qwen3-8B-abliterated using the task_arithmetic method.
  • Context Length: Supports a substantial context window of 32768 tokens, allowing for processing longer inputs and more complex problem descriptions.

Good For

  • Mathematical Reasoning: Ideal for applications requiring robust arithmetic problem-solving and quantitative analysis.
  • Instruction-Based Tasks: Suitable for scenarios where precise adherence to given instructions is critical, especially in a calibrated context.
  • Research and Development: Useful for exploring the effects of task arithmetic merging and AIM calibration on base models for specific domains.

Finetunes
1 models